Performance of Deep Learning Frameworks: Caffe, Deeplearning4j, TensorFlow, Theano, and Torch

This paper presents the comparison of the five deep learning tools in terms of training time and accuracy. The evaluation includes classifying digits from the MNIST data set using a fully connected neural network architecture (FCNN).

With the number of training epochs set to 1, 5, and 10 (and the batch size of 128 images), the report provides comparison across the following parameters:

  • time for training a model
  • classification speed
  • accuracy of classification
  • the amount of code necessary

In addition, you will learn how changing the network “depth” and “width” affects the key evaluation metrics (the data is available for the Tanh and ReLU activation functions).